docs: add dataset snapshot and external run workflow - #903

Adds draft API references and one runnable lifecycle tutorial on top of docs #902.

Scope:

• document immutable dataset-version creation, including explicit case identity, provenance, bounds, permissions, and deliberate non-idempotent retry behavior;
• document external run creation and bounded result ingestion, including pinned evaluator versions, exact plans, receipt replay, atomic batches, and caller-owned execution boundaries;
• add a Python standard-library workflow that creates a snapshot, runs a deterministic retrieval/tool function twice, uploads results, exposes an incomplete comparison, and reports one improvement plus one regression despite equal means;
• update API and dataset guide navigation.

Validation:

• node scripts/audit-links.mjs: 726 pages, zero broken navigation/content links; 29 existing orphan pages;
• npm run build: 1,187 pages built in 20.33 seconds;
• python3 -m py_compile public/cookbook/external-run-comparison.py, Ruff, and git diff --check passed;
• exact tutorial ran through local HTTP with real API-key authentication and disposable PostgreSQL against core lifecycle head 4b30a10796d85bad9390831a87236ffa1ad1eb84: snapshot 201, run/replay 200, incomplete candidate missing one, final means 0.5/0.5 with improved one/regressed one, terminal replay unchanged;
• separate authenticated conflict probe returned 409 for a changed terminal receipt;
• no hosted, model-backed, Observe, browser, SDK, CI release, or deployment qualification is claimed.

The API/SDK endpoints remain draft until their dependent core PRs are reviewed and released. The tutorial uses synthetic callbacks and null trace IDs; it does not imply remote task execution or trace ingestion.